Still holds

Apparently, the difference between "Agile" and "Spec-Driven" development is that in spec-driven development, the specifications are executable. * sigh *

I enjoyed The Odyssey (2026) but I have one historical nitpick. The phrase “Jiminy Christmas” is a minced oath for “Jesus Christ”. As the story is set around 1200 B.C. it is unlikely that Odysseus would have said the phrase even once, let alone the dozens of times he says it in the film

After a bunch of recommendations, finally watched Smile. Disappointing. Hmm.

This is you with Claude Code, this is www.youtube.com/shorts/uqn9r...

My review of my summer holiday reading, @sarahoconnorft.ft.com's "We are not Machines". Strong recommend. constantinides.net/2026/07/16/w...

I see folks talking about a "capability gap" forming btwn teams who use AI & teams who don't There's no evidence of this. If anything, the tech's amplifying capability gaps that *already existed*, making avg & low-performers even worse That's the capability gap you need to address - fundamentals

They can reach him through the ballot box like the rest of us

The other very common consequence of high feedback latency is sunk cost. The further we go down a certain path, the less willing we are to admit we're going the wrong way when reality finally comes a'knockin', and the less likely we are to correct course. codemanship.wordpress.com/2026/07/15/f...

The Luddites loved technology. They owned looms at home and would ‘mod’ them to count threads. Their battle was against worker exploitation. The degradation of their rights, safety and wages. They were against a handful of rich businessmen owning the means of production. Sound familiar?

The front page of your engineering dashboard should show real-world impact. The main reason it doesn't is because you never actually thought about what real-world impact the software's supposed to have. Sure, you're pedaling fast. But you ain't going anywhere.

Learning about "credit assignment depth" in Reinforcement Learning. It's strikingly close to my definition of Feedback Latency - a mirror image, really A model recursively makes a sequence of predictions (e.g., a sequence of tokens) before output's evaluated. Which prediction(s) should be rewarded?

Feedback latency isn’t a measure of time before we get feedback – it’s a measure of unvalidated causality. A team with low output can still easily outlearn teams with much higher output if they have lower feedback latency. codemanship.wordpress.com/2026/07/15/f...

Adding conditions after a contract's been signed is a red flag.

"what used to make someone a 'senior developer' — knowing syntax cold, writing boilerplate fast, remembering every API by heart — is being commoditized in real time"

It's been a while - new video for the Modern Software Engineering channel! On The Horrors Of Finding A New Job

The shamelessness of a political party which is led by a man who just weeks ago called for "pure cold rage" on the streets, shortly before migrants were literally burnt out of their homes, now accusing their opponents of "stochastic terrorism" against them, is quite something

I FUCKING KNEW IT.

Lean and systems thinking has much to say about the impact of batch size on queues, lead times and throughput. Viewed through the lens of Feedback Latency, though, we see a potentially much bigger impact. codemanship.wordpress.com/2026/07/15/f...

Lean and systems thinking has much to say about the impact of batch size on queues, lead times and throughput. Viewed through the lens of Feedback Latency, though, we see a potentially much bigger impact. codemanship.wordpress.com/2026/07/15/f...

A software release is a question, not an answer.

After the upcoming Butlerian Jihad in software development, who will be our Mentats?

And you thought WIP and batch size was all about queues... codemanship.co.uk/sim/feedback...

So much of the thinking around Work In Progress and batch size centres on queues. But that's missing a much bigger picture. Even if they had no effect at all on queues, they have a far worse impact on *learning*

I can only stand in awe of the people who believe you can get a software product design right without any real-world feedback from people using it if you just designed it right in the first place. Never, ever go into a casino with your wallet. Or design software, come to that.

A handy visual aid to understanding why your product's usage stats might not be telling you what you think they're telling you.

Most engineering orgs - and I flatter them using that word - measure productivity by how much *stuff* teams ship. This is to fundamentally mis-characterise software development. It's not a delivery process. We're not Uber Eats. It's a *learning* process. And the biggest factor is feedback latency

"But Jason, what if we could compound errors *faster*?"

Don’t know who to credit for this. But it is how I feel.

In the best design processes, reality gets the deciding vote.

Apologies. We would have shipped the software sooner, but we were too busy guessing what you need.

"The problem with Agile is <goes on to describe something the Manifesto for Agile Software Development explicitly discourages>"

After 17 years running a small business, I've learned to be wary of middle men. Very often it turns out the party they're representing is themselves.

I applied for a job as a canoe instructor and they made me take a swimming test, even after they freely admitted that all their instructors use canoes.

I applied for a job as a canoe instructor and they made me take a swimming test, even after they freely admitted that all their instructors use canoes.

Yet to see much in the way of empiricism in "harness engineering". Seems to be operating mostly on feels. Or should I say "vibes"?

We pay lip service to software releases being experiments. If so, then what is your hypothesis? How will you test it? What is the design of your experiment? Collecting usage stats doesn't count. "They took the pills" isn't meaningful data. How will you know if the medicine's working?

Heck of a week. Huge thanks to the final Code Craft cohort - a real pleasure. Officially on holiday for the first time in 17 years. When I come back, I'll be reconcilng my values with a more corporate landscape. Let's see if I can remember how. Peace out.

Most of what I'm hearing about AI-assisted and agentic software development is the equivalent of arguing that we can make Play Your Cards Right go much faster if we don't wait for the cards to be turned over.

The odds are stacked against error-free long-horizon agentic coding. But even if we *could* create models reliable enough and quality gates fast enough and strong enough to achieve it, there's a strong argument for why we *shouldn't* - REALITY GATES Reality, annoyingly, runs at its own pace.

Do the maths - error-free long-horizon agentic programming requires quality gates so good that I seriously doubt the capability of anybody to even get close. Fols who claim to have cracked it do what I call "silent failing". Errors propagate + compound, and agents just keep building on top of them.

Do the maths - error-free long-horizon agentic programming requires quality gates so good that I seriously doubt the capability of anybody to even get close. Fols who claim to have cracked it do what I call "silent failing". Errors propagate + compound, and agents just keep building on top of them.

This morning's Continuous Delivery exercise is underway. First order of business: get their delivery pipeline up and running with a "Hello, world!" walking skeleton, so they can get the peas from the farmer's field to my table lickety-split.

Should we measure developer experience in years or in feedback?

"But that's just 28 million data samples. Let me counter that with an anecdote" is my new favourite LinkedIn genre

Day #2 of the Code Craft training workshop, and we're going to be focusing on refactoring and design principles.

I don't think "AI harness" is the right term for what folks are trying to do. I think "Hessian crucible" might be more accurate. - attempting to turn base data anyone can access into competitive advantage.

We know from the largest datasets that AI coding slows the majority of dev teams down and makes their software less reliable. And we can explain why. We can even explain why individual developers appear more productive at the same time, and why many teams don't even realise it's happening.

Meanwhile, in Clacton...